UGM-Legal-AI: Question and Answer System for Turkish Customs Legislation

Abstract

The application of LLM models to the legal field has not been successful in generating fluent natural language, source-based, time-valid, and verifiable responses. Traditional RAG pipelines accept fetched texts as direct evidence for response; however, legal accuracy depends not on the text itself but on the legislation in force. This paper aims to present the UGM-Legal-AI system, developed for the application environment of Turkish customs law. The system distinguishes itself from classical RAG by separating the layers of incremental access, distinct legal decision, and supervised LLM generation. The Question Understanding Contract (QUC) combines the parser output, Turkish language compiler, and on-demand semantic enrichment into a single, immutable contract. Sub-layers use only this contract. The agent-based dual-channel pipeline accesses official legislation and corpus in parallel. It implements a critic cycle. It mandates source-based responses only through multi-stage gates, including HS code support contracts. Principles for indexing clauses and paragraphs specific to Turkish legal texts, eight skilled work-time contracts, and a cohort-based production evaluation framework have been defined. Observations on the Silver baseline (v116) show that the canonical sections are consistently retrieved (confidence value 0.7916) and the question understanding path after initial warm-up is reduced to less than 1 second, while access latency is revealed to be the dominant bottleneck. The architecture offers a reusable pattern for domain-specific legal AI where hallucination control and decision traceability are primary requirements.